ISSUE 005
AI didn't remove the work. It moved it downstream.
There is a specific failure I have watched happen and had no word for: someone sends you something that looks completely finished, and you lose an afternoon discovering it isn't. Researchers at Stanford and BetterUp named it, then went and measured what it costs. The number is less interesting than where the cost lands.
An idea worth thinking about
Sloppy work takes effort. This takes none, which is the problem.
BetterUp Labs and Stanford's Social Media Lab coined "workslop" for AI-generated work content that masquerades as good work but lacks the substance to meaningfully advance the task. Slick slides, long reports, code with no context — output that looks credible and leaves the actual thinking undone.
Jeff Hancock, the Stanford communication professor on the study, drew the distinction that makes this new rather than eternal: sloppy work still requires a little effort to produce, and workslop requires none. Bad work has always existed, but it used to cost the person making it something. When production cost falls to nearly zero, the volume changes and so does the etiquette around it.
The mechanism is a transfer, not a loss. The sender genuinely saved time — their productivity gain is real and they can feel it. The receiver pays for it by verifying, correcting, or redoing. Measure at the individual level and AI looks like a triumph. Measure at the level of the whole shop and it can come out negative. In a four-person business, downstream is someone you eat lunch with.
A tool I disappeared into
An open dataset of what people actually ask AI to do, by occupation
The Anthropic Economic Index is a recurring open dataset mapping millions of real conversations to the Department of Labor's O*NET task taxonomy and the Standard Occupational Classification system, published on Hugging Face under a CC-BY licence. Six releases have shipped between February 2025 and June 2026, and from June 2026 it moved to monthly aggregates with country and US state breakdowns.
The variable worth your attention splits every interaction into automation — the user delegates the task — and augmentation, where the user works through it alongside the model. That is the same distinction the workslop research is circling from the other side. Full delegation with no review is how output arrives looking finished and isn't. It's also the split Stanford's labour economists use when they test whether AI exposure is associated with employment declines.
State the obvious limit, because it's a large one: this is Anthropic's data about Anthropic's own product. It describes how people use Claude, not how people use AI. Anyone working exclusively in a different tool is invisible in it. It's used here because it's genuinely open and because independent researchers rely on it, not because it's complete.
Learn something
What happens to an organisation's knowledge when nobody checks
In June 2026 Harvard Business Review published work by Matthias Holweg of Oxford and Thomas Davenport of Babson arguing that the damage from low-quality AI output goes well past individual errors. They call the result knowledge decay: when polished-but-wrong material circulates through documents and systems, colleagues downstream spend time verifying and correcting it, and as errors compound across teams the organisation's collective knowledge base degrades.
It's a useful frame because it explains a thing that otherwise looks like paranoia. The risk isn't one bad report. It's that the bad report gets cited in the next document, which informs a decision, which becomes the thing everyone remembers as true. Small organisations are not immune here; they're arguably more exposed, because there are fewer people positioned to notice.
The practical response both papers converge on is unglamorous: be explicit as a business about when and how AI gets used, and keep a named human accountable for output that leaves the building. Not a policy document. Just a norm that somebody read it.
What we’re seeing
Forty percent of desk workers got workslop last month. Each incident took about two hours to fix.
BetterUp and the Stanford Social Media Lab surveyed 1,150 full-time US desk workers in September 2025. Forty percent said they had received workslop in the previous month, and respondents estimated roughly 15 percent of the content they receive at work qualifies. Each incident took an average of about two hours to resolve.
Converted to money using respondents' own reported salaries, that comes to roughly $186 per employee per month, which the researchers scale to more than $9 million a year for a 10,000-person organisation. I'd treat that last figure as illustrative rather than measured — it's a self-reported time estimate multiplied by a self-reported salary and then extrapolated, and different write-ups of the same study use slightly different percentages to get there.
The findings I'd actually bank are the ones that don't require arithmetic. It flows in every direction: around 40 percent of workslop moves between colleagues, 18 percent goes upward to managers, and 16 percent comes down from managers and executives. This is not a junior-staff problem.
And the interpersonal damage is the part that should worry a small business most. Fifty-three percent of recipients reported being annoyed, 38 percent confused, 22 percent offended. About half said they subsequently viewed the sender as less creative, less capable, less reliable. One respondent described receiving poor work from a supervisor and feeling unable to raise it, so absorbing the rework silently — which is exactly how this stays invisible to the person causing it.
In a business of six people you cannot afford for two of them to quietly reclassify each other as unreliable. That, rather than the dollar figure, is the cost I'd watch for.
One question
Have you ever received something you could tell was AI-written?
What gave it away, and what did you do about it? I'm especially interested in whether anyone actually said something.
More in this pillar as it's written.
— Shana
Sources link to the original publisher, never to a summary.